Research map / Image generation / Autoregressive and latent generation
Infinity: Scaling Bitwise AutoRegressive Modeling for High-Resolution Image Synthesis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Infinity is a bitwise visual autoregressive text-to-image model that uses a bitwise multi-scale quantizer with an infinite-vocabulary classifier and bitwise self-correction to generate high-resolution images from text.
Problem
Autoregressive text-to-image models lag diffusion models in reconstruction fidelity, detail, and quality because index-wise discrete tokens with small vocabularies incur quantization error and fuzzy supervision, and teacher forcing creates train-test discrepancies.
Approach
Replace index-wise visual tokens with bitwise tokens from a binary (BSQ) multi-scale residual quantizer so the vocabulary can be scaled to 2^64, predict each token as d independent bit classifiers (IVC), and randomly flip bits and re-quantize residuals during training (bitwise self-correction) to reduce train-test discrepancy.
Key results
- Infinity-2B reaches GenEval overall 0.73 (with rewriter) and DPG overall 83.5, and ImageReward 0.953 and HPSv2.1 32.05, the top scores in Table 2.
- Scaling the tokenizer vocabulary to 2^64 lowers ImageNet-256 rFID from 1.22 (2^16) to 0.33, versus 0.87 for the SD VAE.
- Generates a 1024x1024 image in 0.8s, 2.6x faster than SD3-Medium.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Scale the discrete tokenizer vocabulary from Vd=216 to Vd=264 (ablation) | Infinity visual tokenizer · ImageNet-256 | rFID: 1.22 -> 0.33 | Helped |
| Use the Infinite-Vocabulary Classifier instead of a conventional index-label classifier (ablation) | Infinity · Text-to-image validation | FID: 5.95 -> 5.32 | Helped |
| Use the Infinite-Vocabulary Classifier instead of a conventional index-label classifier (ablation) | Infinity · Text-to-image validation | ImageReward: 0.79 -> 0.91 | Helped |
| Add bitwise self-correction with re-quantization (ablation) | Infinity · Text-to-image validation, 512×512 | FID: 9.76 -> 3.47 | Helped |
| Add bitwise self-correction with re-quantization (ablation) | Infinity · Text-to-image validation, 512×512 | ImageReward: 0.52 -> 0.76 | Helped |
| Add bitwise self-correction with re-quantization (ablation) | Infinity · Text-to-image validation, 512×512 | HPSv2.1: 29.53 -> 30.71 | Helped |
| Add random bit flipping without re-quantization self-correction (ablation) | Infinity · Text-to-image validation, 512×512 | FID: 9.76 -> 9.69 | No clear effect |
| Add random bit flipping without re-quantization self-correction (ablation) | Infinity · Text-to-image validation, 512×512 | ImageReward: 0.52 -> 0.52 | No clear effect |
| Add random bit flipping without re-quantization self-correction (ablation) | Infinity · Text-to-image validation, 512×512 | HPSv2.1: 29.53 -> 29.20 | Hurt |
| Scale vocabulary from Vd=216 to Vd=232 (ablation) | Infinity 2.2B · Text-to-image generation | evaluation performance | Helped |
Surprising findings
- Small models (125M, 361M) converge faster with a 2^16 vocabulary; the 2^32 vocabulary only wins once the transformer reaches 2.2B parameters.
- Random bit flipping alone without re-quantization gives no FID improvement (9.69 vs 9.76), so the gain comes from re-quantizing the residuals.
Experimental setup
Datasets: LAION COYO OpenImages ImageNet
Benchmarks: GenEval DPG-Bench ImageNet-256 rFID ImageNet-512 rFID
Metrics: GenEval overall DPG overall rFID ImageReward HPSv2.1 VQAScore FID generation latency human preference win rate
Baselines: LlamaGen HART Show-o Emu3 SDXL SD3-Medium PixArt-Sigma FLUX-dev DALL-E 3
Builds on: VAR BSQ LFQ Flan-T5
Closest papers by meaning
- InfinityStar: Unified Spacetime AutoRegressive Modeling for Visual Generation (NeurIPS 2025 · similarity 0.91)
- Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (NeurIPS 2024 · similarity 0.90)
- BitDance: Scaling Autoregressive Generative Models with Binary Tokens (NeurIPS 2026 · similarity 0.89)
- Visual Autoregressive Modeling for Image Super-Resolution (ICML 2025 · similarity 0.89)
- Fundamental Limits of Visual Autoregressive Transformers: Universal Approximation Abilities (ICML 2025 · similarity 0.88)
- Fluid: Scaling Autoregressive Text-to-image Generative Models with Continuous Tokens (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.